The funniest thing about frontier AI is that everyone wants the future until the invoice arrives.
That is why the GLM-5.2 conversation keeps getting louder.
The Atlantic framed GLM-5.2 as China’s answer to AI sticker shock, with developers and founders paying attention because the model is not just cheap, it is useful enough in agentic coding workflows to make people reconsider their default assumptions. TechCrunch is seeing a similar pattern through infrastructure data: DeepSeek has surged in token volume on Vercel’s AI gateway, and Z.ai has moved into a meaningful position too, while expensive frontier models still capture a large share of spend.
That split matters.
It suggests the market is not simply replacing frontier models with cheaper open models. It is sorting work.
Hard discovery work still goes to the expensive frontier layer. Mature production work starts migrating down to cheaper models once the task is understood. The first version needs the sharpest tool. The thousandth repetition needs an efficient one.
That is a very different mental model from the one most people had a year ago.
Cheap models change behavior
When a model is expensive, teams behave carefully. They ration usage, route through approvals, and quietly ask whether the feature is worth the token burn.
When a model is cheap enough, the product shape changes. Agents can run more often. Background jobs become more plausible. Experiments stop feeling like financial sins. Small teams get access to workflows that previously belonged only to companies with absurd compute budgets.
But cheap does not mean free of consequences.
The serious builder still has to ask where the model runs, what data crosses the boundary, what the license allows, how stable the provider is, and whether “good enough” is actually good enough for the failure mode in front of them.
Still, the pressure is real.
The frontier labs can keep owning the top of the capability curve. But if open and lower-cost models keep eating the repetitive production layer, the economics of AI products change fast.
The next war is not only who has the smartest model.
It is who can route intelligence to the cheapest acceptable place without losing trust.
Sources: The Atlantic, TechCrunch, Z.ai